1#version 450
 2
 3#include "generic_head.glsl"
 4#include "types.glsl"
 5
 6#extension GL_EXT_control_flow_attributes : enable
 7#define BLOCK_SIZE 512
 8
 9layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
10
11layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
12layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
13
14shared vec2 sum[BLOCK_SIZE];
15
16void main() {
17    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
18    const uint tid = gl_LocalInvocationID.x;
19
20    sum[tid] = vec2(0.0f, 0.0f);
21
22    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
23        const float xi = float(data_a[row*p.KX + col]);
24        sum[tid].x += xi;
25        sum[tid].y += xi * xi;
26    }
27
28    // sum up partial sums and write back result
29    barrier();
30    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
31        if (tid < s) {
32            sum[tid] += sum[tid + s];
33        }
34        barrier();
35    }
36
37    const float mean = sum[0].x / p.KX;
38    const float var = sum[0].y / p.KX - mean * mean;
39    const float inv_std = inversesqrt(var + p.param1);
40
41    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
42        data_d[row*p.KX + col] = D_TYPE((float(data_a[row*p.KX + col]) - mean) * inv_std);
43    }
44}